We’re doing AI all wrong. Here’s how to get it right | Sasha Luccioni

Episode Summary

Executive Summary: Sasha Luccioni argues that the real AI debate is not whether AI will save or destroy humanity, but whether it will be built sustainably and equitably. She criticizes the “bigger is better” race among major tech firms, highlights the environmental and social costs of large models, and advocates for smaller, task-specific AI, better transparency, and regulation so AI serves people and the planet.

Main Topics: AI’s real problem: sustainability, not hype (Priority: 5/5): Luccioni reframes the AI conversation away from superintelligence fears or utopian promises and toward the immediate harms of resource-intensive AI development. Big AI’s resource-heavy arms race (Priority: 5/5): She criticizes major corporations for building ever-larger models and data centers that consume massive energy, water, and capital while externalizing environmental costs. Small language models as a better alternative (Priority: 5/5): She presents small LLMs as more efficient, cheaper, and often safer, with comparable performance for many tasks and the ability to run on phones or browsers. AI beyond LLMs for climate and public good (Priority: 4/5): She argues that many climate and scientific problems require specialized AI approaches rather than general-purpose chatbots, citing examples in ecology, agriculture, and energy. Transparency and accountability for AI energy use (Priority: 5/5): She describes the AI Energy Score Project as a way to help users compare models by energy efficiency and push companies toward disclosure. Policy and collective action (Priority: 4/5): She calls for laws, incentives, and user pressure to make AI companies accountable and to steer the industry toward sustainable, public-interest outcomes.

Key Arguments: The dominant AI narrative is misleading; the urgent issue is that AI is being built in ways that harm people and the planet. Large general-purpose models often use far more energy than task-specific models, even for simple queries. A recent study cited in the talk found LLMs can use up to 30 times more energy than smaller models for simple questions like identifying Canada’s capital. Small LLMs can deliver strong performance with far less data, compute, and energy, making AI more accessible to startups, academics, nonprofits, and users. Smaller models can improve privacy, cybersecurity, and data sovereignty because they can run locally on devices instead of in massive data centers. AI should be applied selectively: climate, agriculture, biodiversity, and energy systems often need specialized models rather than chatbots. Users currently lack the information needed to choose low-carbon AI tools, so energy/carbon labeling is necessary. Regulation is lagging behind the climate and AI races, so public pressure and policy are needed to force accountability. The industry should learn from the extractive patterns of big oil and avoid repeating them in AI infrastructure and business models.

Data Points: Meta data center scale: Size of Manhattan - Luccioni cites Meta’s planned data center as an example of AI infrastructure expansion. OpenAI Stargate emissions: 3.7 million tons of CO2 equivalents per year - Projected annual emissions for the first phase of OpenAI’s Texas data center. Comparison for Stargate emissions: As much as the whole country of Iceland - Used to contextualize the scale of the projected emissions. XAI gas turbines: 35 turbines - Luccioni says XAI is being sued over air pollution from turbines powering its Colossus data center. Small LLM size: Around 135 million parameters - She describes the smallest small-LM family model as dramatically smaller than traditional LLMs. Relative size vs. DeepSeek: 5,000 times smaller - Comparison used to illustrate the scale difference between small LLMs and a large model. Energy use difference for simple question: Up to 30 times more energy - Her study comparing LLMs to smaller task-specific models for answering simple questions. Training data composition: 60% educational web pages - Hugging Face small LLM training data was curated for quality and lower toxicity/misinformation risk. AI Energy Score sample result: 0.007 watt hours - Energy used by a small model (“Small Alarm”) to answer the capital of Canada question. Energy comparison example: 150 times more energy - DeepSeq’s energy use for the same simple question, compared with Small Alarm. Model evaluation scope: Over 100 open source AI models - Number of models tested in the AI Energy Score Project across text and image tasks.

Pivotal Quotes: "In my opinion, both of these statements are wrong. And what they do is to distract us from the real issue at hand. We're doing AI wrong at the expense of people and the planet." — Sasha Luccioni: Her central thesis reframing the AI debate away from existential hype toward sustainability. "Today, we use AI as if we were turning on all of the lights of the stadium just to find a pair of keys." — Sasha Luccioni: A metaphor illustrating the wastefulness of using massive models for trivial tasks. "A future in which AI serves all of humanity and not just a handful of for-profit tech companies." — Sasha Luccioni: Her closing vision for a more equitable and sustainable AI ecosystem.

Implications: Listeners are urged to demand efficient, transparent AI and to choose tools based on need, not hype. For industry and policymakers, the message is clear: build smaller, greener models, disclose impacts, and regulate AI like a public-interest infrastructure.